Automatic SNR measurement of brain MR images using a deep learning-based approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Quality Assessment

Background:

• Signal-to-noise ratio (SNR) is crucial for MRI quality but traditional measurements are slow and subjective.
• Deep learning presents an opportunity to automate and standardize SNR assessment.

Purpose of the Study:

• To develop and validate a deep learning-based method for automatic SNR measurement using single MRI images.
• To enable observer-independent quantification of MRI image quality.

Main Methods:

• A Pix2Pix deep learning framework with U-Net++ and GAN was trained on brain MRI scans (T1WI, T2WI, FLAIR).
• The model generated signal and noise maps to compute SNR maps pixel-wise.
• Automatic segmentation of whole-brain, white matter (WM), and cerebrospinal fluid (CSF) regions for regional analysis.
• Comparison against the subtraction-map method using SSIM, correlation, and Bland-Altman analyses.

Main Results:

• High agreement with the reference method across all sequences (SSIM: 0.95 ± 0.02).
• Strong correlations (r > 0.86) and low relative errors (<7%) for whole-brain, WM, and CSF SNR values.
• Bland-Altman analysis showed minimal bias and narrow limits of agreement.

Conclusions:

• The developed deep learning method offers automatic, accurate, and observer-independent SNR quantification.
• This tool supports enhanced clinical and research evaluation of MRI image quality.